Abstract
Various defense models have been proposed to resist adversarial attack algorithms, but existing adversarial robustness evaluation methods often overestimate the actual robustness of these models by failing to approach the theoretical lower bound. To address this, we propose the decouple space method, which divides a classifier into non-linear and linear components. By defining a representation space and utilizing iterative Absolute Classification Boundaries Initialization (ACBI), we identify superior attack starting points. Evaluations across nearly 50 widely-used defense models demonstrate that ACBI consistently achieves lower robust accuracy, revealing hidden vulnerabilities in state-of-the-art defenses.
Introduction
Adversarial attacks—small, human-imperceptible perturbations designed to fool neural networks—have led to a continuous arms race between attack and defense strategies. A critical challenge in this field is the accurate evaluation of a model's robustness. Current evaluation methods typically initiate attacks from random points within a constrained neighborhood of the original image. However, such random initialization is often suboptimal; if an attack starts too far from the model's internal decision boundaries, it may fail to find an adversarial example even when one exists nearby. To find the true "lower bound" of robustness, attacks must be initialized as close to the decision boundaries as possible.
Methodology: The Decouple Space Method
The primary innovation of this work is the decoupling of a classifier $f$ into two distinct components:
- The Head ($f_h$): Comprising all layers except the final one, this component performs non-linear feature extraction.
- The Tail ($f_t$): The final fully connected layer, which acts as a linear classifier.
The output of $f_h$ is defined as the representation vector. While decision boundaries in the original high-dimensional image space are complex and non-linear, they become linear in this representation space. These are defined as "Absolute Classification Boundaries."
Absolute Classification Boundaries Initialization (ACBI) leverages this linearity. The process involves:
- Selecting a random starting point.
- Applying gradient descent in the input space to move the corresponding representation vector toward the nearest absolute classification boundary.
- Launching a standard iterative attack (e.g., PGD or FAB) from this optimized starting point.
Experimental Results
We evaluated ACBI on standard benchmarks, including CIFAR-10 and CIFAR-100, across approximately 50 defense models and 8 different architectures. The results consistently show that ACBI uncovers lower robust accuracy than traditional random initialization methods.
Impact of Initialization Iterations
A key finding is the relationship between the number of ACBI initialization steps (init_num) and the resulting robust accuracy. As shown in the figure below, increasing the iterations of the ACBI phase consistently leads to a more effective attack, further reducing the measured robust accuracy across a wide variety of defense models, including DNR, OAAT, and FAT.
For instance, on a ResNet-18 model with standard adversarial training, a standard PGD attack might result in 74.09% robust accuracy. Integrating ACBI (PGD*) causes this figure to drop significantly to 61.48%, demonstrating that the model was previously perceived as more robust than it truly was.
Conclusion
By rethinking the relationship between classifiers and adversarial attacks through space decoupling, we demonstrate that current robustness evaluations are often too optimistic. ACBI provides a modular, "plug-and-play" warm-up phase that can be integrated into existing iterative attacks to more accurately estimate the lower bound of model robustness. Future work will explore refining gradient optimization through the non-linear head to further improve attack efficiency in complex architectures.
Summary updated